Re: [RFC PATCH 00/18] Multi-pack index (MIDX)

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 6/6/2018 4:13 AM, Ævar Arnfjörð Bjarmason wrote:
On Mon, Jan 08 2018, Derrick Stolee wrote:

On 1/7/2018 5:42 PM, Ævar Arnfjörð Bjarmason wrote:
On Sun, Jan 07 2018, Derrick Stolee jotted:

      git log --oneline --raw --parents

Num Packs | Before MIDX | After MIDX |  Rel % | 1 pack %
----------+-------------+------------+--------+----------
          1 |     35.64 s |    35.28 s |  -1.0% |   -1.0%
         24 |     90.81 s |    40.06 s | -55.9% |  +12.4%
        127 |    257.97 s |    42.25 s | -83.6% |  +18.6%

The last column is the relative difference between the MIDX-enabled repo
and the single-pack repo. The goal of the MIDX feature is to present the
ODB as if it was fully repacked, so there is still room for improvement.

Changing the command to

      git log --oneline --raw --parents --abbrev=40

has no observable difference (sub 1% change in all cases). This is likely
due to the repack I used putting commits and trees in a small number of
packfiles so the MRU cache workes very well. On more naturally-created
lists of packfiles, there can be up to 20% improvement on this command.

We are using a version of this patch with an upcoming release of GVFS.
This feature is particularly important in that space since GVFS performs
a "prefetch" step that downloads a pack of commits and trees on a daily
basis. These packfiles are placed in an alternate that is shared by all
enlistments. Some users have 150+ packfiles and the MRU misses and
abbreviation computations are significant. Now, GVFS manages the MIDX file
after adding new prefetch packfiles using the following command:

      git midx --write --update-head --delete-expired --pack-dir=<alt>
(Not a critique of this, just a (stupid) question)

What's the practical use-case for this feature? Since it doesn't help
with --abbrev=40 the speedup is all in the part that ensures we don't
show an ambiguous SHA-1.
The point of including the --abbrev=40 is to point out that object
lookups do not get slower with the MIDX feature. Using these "git log"
options is a good way to balance object lookups and abbreviations with
object parsing and diff machinery.[...]
[snip]

[...]And while the public data shape I shared did not show a
difference, our private testing of the Windows repository did show a
valuable improvement when isolating to object lookups and ignoring
abbreviation calculations.
Replying to this old thread since I see you're prepearing the MIDX for
submission again and this seemed like the best venue.

You're really good at tracking new commits in my GitHub page ;)


Your WIP branch (github.com/git/derrickstolee/midx/upstream) still only
references the speedups in abbreviation calculations, but here you
allude to other improvements. It would be very nice to have some summary
of that in docs / commit messages when you submit this.

The new version is essentially a complete rewrite of the feature, since we learned a lot about how to add a new data store from the commit-graph series. The design document [1] refers to some of the immediate benefits and future benefits. Some of these future benefits were discussed at the contributor's summit [2].

[1] https://github.com/derrickstolee/git/blob/midx/upstream/Documentation/technical/midx.txt

[2] https://public-inbox.org/git/alpine.DEB.2.20.1803091557510.23109@alexmv-linux/
    Git Merge 2018 Contributor's summit notes (includes discussion of MIDX)


I've been meaning to get around to submitting something like I mentioned
in https://public-inbox.org/git/87efn0bkls.fsf@xxxxxxxxxxxxxxxxxxx/
i.e. a way to expand the abbrev mode to not check disambiguations, which
would look something like:

     core.abbrev = 20
     core.validateAbbrev = false

Or:

     core.abbrev = +2
     core.validateAbbrev = false

So, using the example from the above referenced E-Mail +2 would make
linux.git emit hashes of 14 characters, without any abbreviation
checking (just trusting in statistics to work in your favor).

As noted by JS in this thread that wouldn't be acceptable for your
use-case, but there's plenty of people (including me) who'd appreciate
the speedup without being a 100% sure we're emitting unambiguous hashes,
since that trade-off is better than time spent generating another index
on-disk. So I see it as a complimentary & orthogonal feature.

But with that implemented I wouldn't get any benefit from things that
use the MIDX that aren't abbreviations, so what are those?

The MIDX is built for handling many packfiles. As opposed to the commit-graph feature, your repo needs to be _really_big_ to need the MIDX. Most just repack into one packfile on a regular basis.

One case for vanilla Git: we've heard from lots of customers disabling gc.auto in their build machines because they didn't want to wait for a repack/gc after a fetch and before a build. Then, they end up in a many-pack situation because they never scheduled time for that repack/gc.

For GVFS, we virtualize the repo, so objects can be downloaded from the server on-demand. Since round-trips to the server are expensive, we have the notion of a "prefetch pack". Our cache servers precompute packfiles containing all of the commits and trees introduced to the server that day. On a 'git fetch' command, GVFS intercepts via a hook and downloads the list of daily prefetch packs since the last prefetch. These are placed in an alternate (the "shared object cache"), so multiple enlistments can point to the same, fixed list of packs. The cache servers fold these packs together after 30 days, so users on a fresh clone start with 31 packs and users who go on vacation for a month get at most 31 packs in a single download.

The performance gains from the MIDX for abbrevations are the easiest to demonstrate. It's not hard to see why the single binary search to find the two closest OIDs is better than iterating through all of the packfiles and running a binary search on each.

The other thing that the MIDX does right now is be a single source for object lookups. This is harder to demonstrate the benefit in Git on its own, because using `git repack -adfF --max-pack-size=128m` repacks in a "good" way. This repack actually makes the MRU packfile list really effective.

One reason the MIDX is important for object lookups in the GVFS case is two-fold:

1. The time-based organization of the prefetch packs makes the trees necessary for a 'checkout' spread across many packfiles. Even with the working directory virtualization, every 'checkout' needs to recompute the .git/index, and that requires walking lots of trees.

2. When we are looking for a blob that is not in the local cache, we need to check that the blob is on-disk before requesting it from the server. In a many-pack scenario, those object misses are as expensive as the abbreviation case. This may affect partial clone, so I'll talk with Jeff H about that situation.

I expect to send my v1 today. I'm just trying to create a synthetic environment that demonstrates this benefit to object lookups using the Linux repository.

Thanks,
-Stolee



[Index of Archives]     [Linux Kernel Development]     [Gcc Help]     [IETF Annouce]     [DCCP]     [Netdev]     [Networking]     [Security]     [V4L]     [Bugtraq]     [Yosemite]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux SCSI]     [Fedora Users]

  Powered by Linux